It makes more sense to move all scan stuff in the same place.
Also, we don't really need to duplicate the uses_primid field
for each stages.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
sync_files are in linux since 4.7, while the amdgpu fence_to_handle
ioctl is only in 4.15.
In particular we don't need it for sync_file in radv, because
everything happens via syncobjs, which got support earlier than
fence_to_handle.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
... to set_vs_specific_input_locs().
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The idea is to clean up the add arguments logic.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
... to declare_global_input_sgprs().
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
... to declare_vs_specific_inputs_sgprs().
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
It's a no-op anyway but it looked strange to me, remove it.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
These are just taken from amdvlk, we probably knew these already,
but may as well port them now.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is not because the vertex stage needs some push constants
that other stages need them too. This should reduce the number
of loaded SGPRs in some situations.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
pipeline->layout can't be NULL now.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This wasn't calculating the correct value, this along with
a nir patch fixes a regression in:
dEQP-VK.tessellation.shader_input_output.barrier
Fixes: 043d14db30 (ac/nir: don't write tcs outputs to LDS that aren't read back.)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
And move the comment to amd/common.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Use 16_ABGR instead of 32_ABGR if Z isn't written.
Ported from RadeonSI.
No CTS regressions on Polaris.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
ac_shader_util.c will contain shader helpers for RadeonSI
and RADV.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Use a boolean instead because the number of needed SGPRs
is always 3.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The number of grid components is always 3 when gl_NumWorkGroups
is declared, because it relies on the number of components of
nir_instrinsic_load_num_work_groups.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Without this we get the error "FPExt only operates on FP" when
converting the following:
vec1 32 ssa_5 = b2f ssa_4
vec1 64 ssa_6 = f2f64 ssa_5
Which results in:
%44 = and i32 %43, 1065353216
%45 = fpext i32 %44 to double
With this patch we now get:
%44 = and i32 %43, 1065353216
%45 = bitcast i32 %44 to float
%46 = fpext float %45 to double
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
To support the reindex intrinsic, we need the result to be
something on which we can adjust the index/address.
Since it is all within a basic block, the compiler should be
able to merge any extra loads.
v2: Change visit_get_buffer_size too.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This patch is ported from RadeonSI and it has two effects.
It fixes a rendering issue which affects F1 2017 and Dawn
of War 3 (Vega only) because LLVM was ending up by generating
the new v_mad_mix_{hi,lo} instructions which appear to be
buggy in some way. Not sure if Mesa is generating something
wrong or if the issue is in LLVM only. Anyway, that explains why
the DOW3 issue can't be reproduced with GL on Vega.
It also improves performance because v_cvt_pkrtz_f16 is faster,
and because I guess the rounding mode behaviour is similar between
GL and VK, we can use it. About performance, it improves Talos
by +3/4% but I don't see any other impacts.
No CTS regressions on Polaris.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
V2: make use of driver_location and don't expose NIR to the ABI.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
v4: - Ensure inc_amd_common defined when radeonsi is disabled (needed by
r600)
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Tested-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Fix a bunch of labels indicating when registers were added/removed
and normalize the SI-class GRBM_GFX_INDEX.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The next commit will reduce the size even more.
v2: typecast to uint64_t manually
v3: add more typecasts, add asserts
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
If the TCS doesn't read back the outputs, no need to store them
to LDS in the first place. (except for tess factors).
This seems to give about 50fps (3290->3330) with tessellation demo.
I haven't tested if it impacts DoW3 at all.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is derived from tgsi/radeonsi code from the GLSL intrinsics.
This should pre-fix radv for the upcoming spirv patches.
v2: actually use wait_cnt, sleep deprived dad time! (Bas)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This uses C++11 initializer lists.
I just overwrote all Mesa files with internal addrlib and discarded
hunks that we should probably keep, but I might have missed something.
The code depending on ADDR_AM_BUILD is removed. We can add it back next
time if needed.
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
I was hacking something stupid in doom, and hit an assert for the bitcast
following this, it definitely looks like this should be the number of 32-bit
components, not the instr level ones.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
These are produced by nir_lower_bitmap(), adding the missing derefence
would cause other issues that need to be hacked around such as
skipping sampler lowering and uniform location assignment, so this
change seems the correct way to go.
Fixes 194 piglit crashes on radeonsi using NIR.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
It's inaccurate. Instead, see the copyright and use "git log" and
"git blame" to know the authorship.
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This was duplicated between both drivers, share here.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is needed for RADV to support explicit component packing.
This is also required to use the new NIR component splitting /
packing passes.
V2:
- add commponent packing support for interpolate_at* intrinsics
- improve store packing support when not all varyings are scalar
as spotted by Bas the store source was incorrectly offset.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This restores performance for the drirc workaround, i.e.
KILL_IF does:
visible = src0 >= 0;
kill_flag &= visible; // accumulate kills
amdgcn_kill(wqm_vote(visible)); // kill fully dead quads only
And all helper pixels are killed at the end of the shader:
amdgcn_kill(kill_flag);
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Vulkan CTS does not expect the value to be clamped (at least for D32),
and it makes a differences even though depth is in [0,1], due
to strict inequalities.
I couldn't find anything in the Vulkan spec about this, but the test
seemed to be copied from GL tests and the GL spec only specifies
clamping for fixed point formats. Hence I expect radeonsi to run into
this at some point as well, but given that they still have a usecase
with the Z16->Z32 promotion, I'll leave that for someone else to clean
up.
This at least fixes radv dEQP-VK.texture.shadow.* on VI.
Fixes: 0f9e32519b 'ac/nir: clamp shadow texture comparison value on VI'
Reviewed-by: Dave Airlie <airlied@redhat.com>
Otherwise we just need to write them to the tf ring.
this seems to improve the tessellation demo on Bonarie
~2190->~2230 fps
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
They don't take a single wave anymore and we need the barriers.
Fixes: 6bc42855f9 'radv: enable GS on GFX9'
Reviewed-by: Dave Airlie <airlied@redhat.com>
For Vulkan SPIR-V the spec states
fma() Inherited from OpFMul followed by OpFAdd.
Matt says the backend will do the right thing depending on the
hardware being compiled for, if you use the fmuladd intrinsic.
Using the Mad Max pts test, on high settings at 4K:
CHP: 55->60
HGDD: 46->50
LM: 55->60
No change on Stronghold.
Thanks to Feral for spending the time to track this down.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.2" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The hardware usually does this automatically. However, we upgrade
depth to Z32_FLOAT to enable TC-compatible HTILE, which means the
hardware no longer clamps the comparison value for us.
The only way to tell in the shader whether a clamp is required
seems to be to communicate an additional bit in the descriptor
table. While VI has some unused bits in the resource descriptor,
those bits have unfortunately all been used in gfx9. So we use
an unused bit in the sampler state instead.
Fixes dEQP-GLES3.functional.texture.shadow.2d.linear.equal_depth_component32f
and many other tests in dEQP-GLES3.functional.texture.shadow.*
Fixes: d4d9ec55c5 ("radeonsi: implement TC-compatible HTILE")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Avoid a v_cndmask: the absolute value is free due to input modifiers.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Fix the custom cube coord selection sequence to be identical to
the hardware v_cubesc/tc and OpenGL spec. Affects texture sampling
with user-provided derivatives.
Fixes dEQP-GLES3.functional.shaders.texture_functions.texturegrad.*
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Redundant with the recently added ac_llvm_context::chip_class.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This builds, installs, and has been tested on a r290x (Hawaii) with the Vulkan
CTS. It dies horribly in a fire at the same point for the meson build as the
autotools build.
v2: - enable radv by default
- add shader cache support and enforce that it's built for radv
v3: - Fix typo in meson_options (Nicholas)
- strip trailing 'svn' from llvm version before setting the version
preprocessor flag (Bas)
- Check for LLVM module requirements
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
If we don't have a depth piece, we don't get a correct
swizzle mode and we hit an assert in addrlib.
In case of no depth get the preferrred swizzle mode for
stencil alone.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Cc: "17.2" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The NIR-to-LLVM pass already does this; now the same fix covers
radeonsi as well.
Fixes various tests of
dEQP-GLES31.functional.texture.filtering.cube_array.combinations.*
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Only on GFX9 we implement them as 2D images.
This fixes:
dEQP-VK.image.image_size.1d_array.readonly_12x34
dEQP-VK.image.image_size.1d_array.readonly_1x1
dEQP-VK.image.image_size.1d_array.readonly_32x32
dEQP-VK.image.image_size.1d_array.readonly_7x1
dEQP-VK.image.image_size.1d_array.readonly_writeonly_12x34
dEQP-VK.image.image_size.1d_array.readonly_writeonly_1x1
dEQP-VK.image.image_size.1d_array.readonly_writeonly_32x32
dEQP-VK.image.image_size.1d_array.readonly_writeonly_7x1
dEQP-VK.image.image_size.1d_array.writeonly_12x34
dEQP-VK.image.image_size.1d_array.writeonly_1x1
dEQP-VK.image.image_size.1d_array.writeonly_32x32
dEQP-VK.image.image_size.1d_array.writeonly_7x1
Fixes: 1bcb953e16 "radv: handle GFX9 1D textures"
Reviewed-by: Dave Airlie <airlied@redhat.com>
This will allow us to use it from radv.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
For the comp_swap case this was overflowing and crashing
sometimes.
Fixes:
dEQP-VK.image.atomic_operations.compare_exchange.*
Cc: "17.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
As GFX9 can't handle 1D depth textures, radeonsi and
apparantly pro just update all 1D textures to 2D,
and work around it.
This ports the workarounds from radeonsi.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.2" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
There were some overlapping changes in gfx9 especially in the CB/DB
blocks which made register dumps rather misleading.
The split is along the lines of the header files, so we'll print VI-only
fields on SI and CI, for example, but we won't print GFX9 fields on
SI/CI/VI, and we won't print SI/CI/VI fields on GFX9.
Acked-by: Marek Olšák <marek.olsak@amd.com>
Automatically re-use table entries like StringTable and IntTable do.
This allows us to get rid of the "fields_owner" logic, and simplifies
the next change.
Acked-by: Marek Olšák <marek.olsak@amd.com>
The VI sizing only applies to VI.
This fixes:
dEQP-VK.image.image_size.buffer.*
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This reduces the size from 96 to 80 bytes but putting all the
32-bit sizes at the start.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
For radv, in order to report VM faults when detected.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
location is never set to INTERP_SAMPLE, and Nicolai comments:
"... that part is misleading. location refers to the base location, not
the final location of the sample, and it can never be INTERP_SAMPLE."
Suggested-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
These are likely false positives, but are also annoying because they
show up on every "make install", which causes ac_nir_to_llvm to be
rebuilt here. Initializing those variables to NULL should be harmless
even when unnecessary.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
amd/common/ and amd/vulkan/ are using tabs for indent, which doesn't
match the settings in root .editorconfig, so let's override.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Acked-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This reverts commit fc99cb3c9e.
"The performance went down from 64.7 to 51.4 fps in Valley and from 30.8 to
25.1 fps in Heaven on Radeon HD 7970. Other games seem to have also a 10-25%
performance decrease."
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102429
It looks like we can't use the raster config values from the kernel.
Similar to e09d04cd56 "radeonsi: use util_strchrnul() to fix android build error"
Android Bionic does not support strchrnul() string function,
gallium auxiliary util/u_string.h provides util_strchrnul()
This change avoids the following warning and error:
external/mesa/src/amd/common/ac_debug.c:501:15: warning: implicit declaration of function 'strchrnul' is invalid in C99
char *end = strchrnul(out, '\n');
^
external/mesa/src/amd/common/ac_debug.c:501:9: error: incompatible integer to pointer conversion initializing 'char *' with an expression of type 'int'
char *end = strchrnul(out, '\n');
^ ~~~~~~~~~~~~~~~~~~~~
1 warning and 1 error generated.
Fixes: c2c3912410 "ac/debug: annotate IB dumps with the raw values"
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Rob Herring <robh@kernel.org>
It uses an user SGPR to pass the view index to the shaders, except
for the fragment shader where we use layer=view (which comes in
handy when we want to do the NV ext that allows us to execute pre-FS
stages once instead of per view).
Reviewed-by: Dave Airlie <airlied@redhat.com>
The int32->float semantic conversion got dropped in a testcase,
because the src was already float. On closer inspection I decided
to add a few more casts for integer op operands to be safe too.
Cc: 17.2 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Seems like we actually enabled it already, but did not implement
the shader part. With this patch we do.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
GFX9 moved where the viewport index export goes.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.2" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Not required anymore.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.2" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
UE4Editor has this issue.
This commit prevents hangs (release build) or assertion failures (debug
build). It doesn't fix the editor, but catastrophic scenarios are
prevented.
Cc: 17.1 17.2 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
The .f32 was already getting added by emit_intrin_2f_param(). Noticed
when enabling LLVM module verification.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This just copies the code from the -pro shaders,
and fixes the tests on CIK.
With this CIK passes the same set of conformance
tests as VI.
Fixes: 83e58b03 (radv: flush f32->f16 conversion denormals to zero. (v2))
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We need vulkan and gl to produce the same UUIDs. Therefore we should
keep the mechanism to compute these in a common location to guarantee
they are updated in lockstep.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Note that dcc_alignment = pipe_interleave_bytes * num_pipes * num_banks,
which is greater than the previous open-coded alignment.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The tile swizzle computation was done after the whole miptree was computed,
but that was too late, because at that point AddrSurfInfoOut contained
information about the smallest miplevel, which is never 2D-tiled.
The correct way is to do the computation before the second level is computed.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
There is a firmware regression that causes failures. Work around it by
using the compute shader for query_buffer_objects to summarize the query
results.
v2: rename to PREDICATION_OP_BOOL64 (consistent with sid.h)
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
LLVM complained about passing an i32 to a float clamp.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Fixes: 0f9e32519b "ac/nir: clamp shadow texture comparison value on VI"
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This looks like it's supported since llvm 3.9 at least,
so switch over radeonsi and radv to using it, -pro also
uses this. We can now drop creating lds for these operations
as the ds_swizzle operation doesn't actually write to lds at all.
Acked-by: Marek Olšák <marek.olsak@amd.com>
(stable requested due to fixing radv CIK conformance tests)
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
This makes it match radeonsi. The LLVM backend itself will emit the
correct instruction, but LLVM might do incorrect optimizations since it
thinks the output is undefined when the input is 0, even though it's not
supposed to be. We really need a new intrinsic, or for the backend to
become smarter and recognize this pattern.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Bas Nieuwenhuizen <basni@google.com>
This fixes:
dEQP-VK.api.copy_and_blit.core.blit_image.all_formats.*
for a2r10g10b10 formats as destination on SI/CIK hardware.
This adds support to the meta program for emitting 10-bit
outputs, and adds 10-bit support to the fragment shader key.
It also only does the int8/10 on SI/CIK.
Fixes: f4e499ec7 (radv: add initial non-conformant radv vulkan driver)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Needed for TC-compatible HTILE in radeonsi for test cases like
piglit spec/arb_texture_rg/execution/fs-shadow2d-red-01.shader_test
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This simplifies a bunch of places that no longer need special treatment
of value_count == 1. We rely on LLVM to optimize away the 1-element vector
types.
This fixes a bunch of bugs where 1-element arrays are indexed indirectly.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The scanning phase sets the driver_location, because it is part of the
ABI: radeonsi does the assignment differently.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The eventual goal is to hide all radv-specific details behind
ac_nir_context::abi, so that the NIR->LLVM code can be re-used by
radeonsi.
During development, we live with a partial split, where some of the
NIR->LLVM code still relies on linking back to the nir_to_llvm_context
(which should ultimately be renamed to reflect that it's radv-specific).
The idea is to get rid of these backlinks over time.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This ports 72e46c988 to radv.
radeonsi: apply a TC L1 write corruption workaround for SI
Fixes: f4e499ec7 (radv: add initial non-conformant radv vulkan driver)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This ports: da7453666a
radeonsi: don't apply the Z export bug workaround to Hainan
to radv.
Just noticed in passing.
Fixes: f4e499ec7 (radv: add initial non-conformant radv vulkan driver)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Translate the NIR variables directly to LLVM instead of lowering to a
TGSI-style giant array of vec4's and then back to a variable. This
should fix indirect dereferences, make shared variables more tightly
packed, and make LLVM's alias analysis more precise. This should fix an
upcoming Feral title, which has a compute shader that was failing to
compile because the extra padding made us run out of LDS space.
v2: Combine the previous two patches into one, only use this for shared
variables for now until LLVM becomes smarter.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Tested-by: Alex Smith <asmith@feralinteractive.com>
During bring-up, this is often 0. Prevent automatic disablement of
ARB_timer_query and demotion of the OpenGL version to 3.2 by setting
a non-zero frequency. Print an error message instead.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This fixes the misspelling of ALIGNMENTS in addrlib.
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This patch uses addrlib to workout the tile swizzles according
to the surface index. It seems to produce the same values as
amdgpu-pro for the deferred test.
v2: don't apply swizzle to CMASK. the eg docs don't mention
it, and we clearly don't align cmask for that.
v3: disable surf index for dedicated images, as these will
most likely be shared, and I don't think the metadata has
space for this info in it yet.
v4: update for shareable images, rename combined_swizzle
to tile_swizzle
This gets the deferred demo from 730->950fps on my rx480.
(dcc cmask elim predication patches get it further)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Since radv uses compute rings and we can't know when we are setting
up the shaders what ring they are to be used on, we should just use
the default xnack setting. This may be suboptimal in some places,
but if we hit a problem, we likely should try and address this
between llvm and mesa.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Rather than using 64k, use what addrlib returns as the base
alignment for vulkan allocations.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The NIR parameters are ordered "compare, data", matching GLSL, but both
the image and buffer LLVM intrinsics take them the other way around.
This is already handled correctly for SSBO atomics.
Signed-off-by: Alex Smith <asmith@feralinteractive.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Fixes: f4e499ec79 "radv: add initial non-conformant radv vulkan driver"
Just more moving code around before adding things to it.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This just modifies the API to make it easier to add other flags
to target machine creation.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
According to Nicolai the SX can already start work when all
the position exports are done, so do those first.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
NIR always makes the shift amount 32 bits, but LLVM asserts if the two
sources aren't the same type. Zero-extend the shift amount to make LLVM
happy.
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
We implement the split opcodes, and tell NIR to lower the original ones.
The lowering to LLVM is a little more complicated, but NIR can optimize
the split ones a little better, and some NIR lowering passes that we
might want to use (particularly for doubles) emit the split ones.
This should fix pack/unpackDouble2x32, which seems like a bug since when
we enabled the Float64 capability. It will also fix pack/unpackInt2x32
when we enable the Int64 capability.
Fixes: 798ae37c ("radv: Enable Float64 support.")
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
We apparently still used v16i8 ....
As radeonsi doesn't use it with LLVM version checks I don't think
we need them either.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
The buffer intrinsics should be used instead of the image ones.
Signed-off-by: Alex Smith <asmith@feralinteractive.com>
Cc: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
When using cmpswap on an image, it was being trunctated to
lvm.amdgcn.image.atomic.cmpswa, with the coords type missing entirely.
v2: Add stable CC
CC: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Booleans in NIR are ~0 for true, b2i returns 0/1.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
LLVM has required an i1 here for a long time. llvm.ctlz.* was fixed in
commit edd23e0606 ("ac/llvm: fix various findMSB bugs").
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Noticed in passing.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Just noticed in passing.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The parses skips the line if it contains parentheses.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
The very last entry in the sid_strings_offsets table ended up missing,
leading to out-of-bounds reads and potential crashes.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The commit did not add the relevant includes - in particular
stdint.h and stdbool.h for the respective standard types.
At the same time, the amdgpu_device_handle typedef redeclaration was
off.
Fixes: 81945ded0d ("ac: remove amdgpu.h dependency")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101471
Cc: Mark Janes <mark.a.janes@intel.com>
Cc: Gregor Münch <gr.muench@gmail.com>
Reported-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reported-by: Mark Janes <mark.a.janes@intel.com>
Reported-by: Gregor Münch <gr.muench@gmail.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Add a couple of forward declarations and drop the amdgpu.h requirement.
With this we can build the r300 and r600 drivers without the need for
amdgpu.
v2:
- Add amdgpu.h include in the C file (Marek)
- Add a comment about pre C11 typedef redeclaration warning (Eric)
Cc: Nicolai Hähnle <nicolai.haehnle@amd.com>
Cc: Marek Olšák <marek.olsak@amd.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101189
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
has_hw_decode is assigned twice.
Pointed out by coverity.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Also solve "outinfo may be used uninitialized" warning by putting in an
unreachable().
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Most functions are only inspecting nir, so nir related arguments can be
marked const. Some more can be done if/when some nir changes are
accepted.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This just moves this code in here to it's cleaner.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Instead of having the fragile code to do a second pass, just
give the pointers you want params in to the initial code,
then call a later pass to assign them.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Just pass a pointer and increment inside the function,
makes the code less error prone.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Simple search for a backslash followed by two newlines.
If one of the newlines were to be removed, this would cause issues, so
let's just remove these trailing backslashes.
Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
We always compute HTILE size using addrlib, even when not TC compatible.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlied <airlied@redhat.com>
This just adds the strings and includes the gfx9 register defs
in some files that we need them in.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is mostly mechanical changes of renaming types and introducing
"legacy" everywhere.
It doesn't use the ac_surface computation functions yet.
Reviewed-by: Dave Airlie <airlied@redhat.com>
This ports: 55445ff189 from radeonsi
radeonsi: tell LLVM not to remove s_barrier instructions
LLVM 5.0 removes s_barrier instructions if the max-work-group-size
attribute is not set. What a surprise.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is needed to add the max workgroup size attribute.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
BOs larger than the minimum fragment size should have their VA
alignet to at least the fragment size for optimal performance.
v2: drop unused leftover from initial implementation
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Commit e1af20f18a changed the shader_info
from being embedded into being just a pointer. The idea was that
sharing the shader_info between NIR and GLSL would be easier if it were
a pointer pointing to the same shader_info struct. This, however, has
caused a few problems:
1) There are many things which generate NIR without GLSL. This means
we have to support both NIR shaders which come from GLSL and ones
that don't and need to have an info elsewhere.
2) The solution to (1) raises all sorts of ownership issues which have
to be resolved with ralloc_parent checks.
3) Ever since 00620782c9, we've been
using nir_gather_info to fill out the final shader_info. Thanks to
cloning and the above ownership issues, the nir_shader::info may not
point back to the gl_shader anymore and so we have to do a copy of
the shader_info from NIR back to GLSL anyway.
All of these issues go away if we just embed the shader_info in the
nir_shader. There's a little downside of having to copy it back after
calling nir_gather_info but, as explained above, we have to do that
anyway.
Acked-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The renumbering code didn't take into account that multiple VS exports
can have the same PARAM index. This also significantly simplifies
the renumbering. Thankfully, we have piglits for this:
spec@arb_gpu_shader5@arb_gpu_shader5-interpolateatcentroid-packing
spec@glsl-1.50@execution@interface-blocks-complex-vs-fs
Reported by Michel Dänzer.
Fixes: b08715499e ("ac: eliminate duplicated VS exports")
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This just adds the chip in the right places.
We don't set the partial_vs_wave workaround, as radeonsi
doesn't, but have to confirm it's not required.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
LLVM 3.8:
- had broken indirect resource indexing
- didn't have scratch coalescing
- was the last user of problematic v16i8
- only supported OpenGL 4.1
This leaves us with LLVM 3.9 and LLVM 4.0 support for Mesa 17.2.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Timothy Arceri
Samuel Pitoiset
Bas Nieuwenhuizen
Marek Olšák
Dave Airlie
Dylan Baker
Nicolai Hähnle
Boyuan Zhang
Matthew Nicholls
Eric Anholt
Andres Rodriguez
Jason Ekstrand
Connor Abbott
Grazvydas Ignotas
Mauro Rossi
Alex Smith
James Legg
Emil Velikov
Eric Engestrom
Leo Liu
Christian König